Logbook
Logbook: Extremely Long-form Audio Event Understanding
A benchmark for hour-scale audio understanding requiring gap-free segmentation with event labels and descriptions from recordings ranging from ten minutes to six days.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info logbook
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Gemma 4 | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Gemma 4 | 0.2 | cascaded (msclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5.6 | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Gemini 3.5 | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen3 Max (2025-09-23) | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Llama 3.3 | 0.2 | cascaded (msclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen2.5 Max | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| GPT-5.6 | 0.2 | cascaded (msclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Gemini 3.5 | 0.2 | cascaded (msclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Llama 3.3 | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| Qwen2.5 Max | 0.2 | cascaded (msclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
| OLMo 3.1 | 0.2 | cascaded (enclap) | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
Curated from Ego4D, EgoLife, and SINS datasets, Logbook features continuous audio recordings up to six days long. Systems must predict non-overlapping segments with activity labels and free-text descriptions.
Evaluation uses Frame-level Accuracy (fAcc) for segmentation/labeling and Description F1 (dF1) for text quality. Best systems remain below human reference, with end-to-end models often outperforming cascaded ones despite degradation over long contexts.